acceptodds
Withdrawn from ICLR 2027

Riemannian Optimization Of Variational Quantum Algorithms

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Quantum Natural Gradient (QNG) is widely considered the theoretically optimal optimizer for variational quantum circuits. However, QNG can become unstable as variational quantum circuits become deeper and noisier, yet the geometric origin of this instability remains poorly understood. We uncover a noise-induced Riemannian spike (NIRS): under local Markovian noise, the pullback metric can vanish exponentially faster than the Euclidean gradient, causing metric inversion to amplify the natural-gradient update exponentially with circuit depth. This reveals a previously overlooked failure mechanism in which noise can destabilize optimization through metric inversion, rather than through gradient suppression alone. In parallel, we introduce Dually Flat Stratified Quantum Natural Gradient (DFS-QNG), a structured Riemannian optimizer that exploits commuting parameter strata. By restricting each update to a commuting submanifold, DFS-QNG obtains exact blockwise geodesic updates while replacing costly global metric inversion with smaller blockwise systems. We further establish geodesic smoothness and convergence guarantees for the resulting Riemannian optimization framework. Experiments on noisy and noiseless VQE and QAOA circuits demonstrate improved stability and faster convergence than standard QNG and Adam, together with substantially reduced metric computation.

Primary area:
optimization
Keywords:
Riemannian Optimization, Information Geometry, Quantum ML

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_tgLpOpenReview ↗
Rating
1
Confidence
4
Informativeness
3
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the stability and computational cost of quantum natural gradient (QNG) optimization for variational quantum algorithms. Its main contributions are a geodesic smoothness analysis with convergence guarantees, a characterization of noise-induced instability caused by the metric shrinking faster than the gradient, and DFS-QNG, a method that updates mutually commuting parameter blocks using smaller metric matrices. Under the stated assumptions, these block updates follow exact geodesics within their restricted submanifolds. Validation consists of VQE and QAOA experiments on 4–10 qubits, covering noiseless and depolarizing-noise settings with five runs per configuration. The reported results show improved optimization stability and lower cumulative energy error compared with standard QNG and Adam. The method reduces metric computation but does not eliminate the underlying noise-induced instability

Critical strengths

The manuscript investigates Riemannian optimization for variational quantum algorithms in both ideal and noisy settings, addressing a conceptually interesting question: how the geometry of the quantum state manifold affects optimization when noise suppresses both gradients and the metric. Its attempt to connect geometric smoothness, noise-induced behavior, and convergence guarantees is a strength. The proposed DFS-QNG framework is also potentially useful: by organizing updates into smaller parameter blocks with commuting generators, it seeks to reduce the cost of metric estimation and inversion while retaining some geometric information. This provides a concrete algorithmic direction beyond the theoretical analysis. However, the significance of these contributions depends on the validity of the convergence arguments and on whether the computational savings translate into improved performance once measurement costs and the effects of blockwise updates are accounted for.

Critical weaknesses

A substantive mathematical error undermines the proof of Theorem 2(i). Immediately below Equation (37), the authors use 1/ϵ=O(q−2a)1/\epsilon=O(q^{-2a}) in the regime ϵ≪q2a\epsilon\ll q^{2a}, although this assumption does not justify the claimed upper bound. If ϵ/q2a→0\epsilon/q^{2a}\to0, then (1/ϵ)/q−2a→∞(1/\epsilon)/q^{-2a}\to\infty, directly contradicting that estimate. Consequently, bounding the inverse regularized metric by 1/ϵ1/\epsilon does not establish the asserted O(q−2a)O(q^{-2a}) scaling. More generally, an upper bound on the metric’s largest eigenvalue cannot provide the lower bound on its smallest eigenvalue needed to control the inverse. This is consequential because the argument for bounded Levi–Civita connection coefficients relies on such inverse-metric control. It is therefore a gap in the proof rather than a typographical error. Repair would require an appropriate spectral lower bound, together with the corresponding derivative estimates, or a different argument. Until this is supplied, the claimed geometric stability in Theorem 2(i) remains unestablished.

Questions for authors

Do you have an alternative proof of Theorem 2(i)?

Reviewer_TZDuOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper has three parts. First, it proves that the variational quantum manifold with a Tikhonov-regularized pullback Fubini-Study metric is geodesically smooth, which yields an O(L_R over delta) rate for Riemannian gradient descent. Second, it argues that under local depolarizing noise the pullback metric contracts as the square of the state contraction rate while the Euclidean gradient contracts linearly, so the natural gradient grows with depth under weak regularization. The authors call this a noise-induced Riemannian spike and contrast it with the noise-induced barren plateau. Third, it introduces DFS-QNG, a block-coordinate version of quantum natural gradient in which each block contains mutually commuting generators. The claim is that within such a block the Christoffel symbols vanish, the Euler step is an exact geodesic, and the metric inversion is a small well-conditioned system. Experiments report normalized area above optimal for VQE on a transverse-field Ising model and for QAOA on ring and 3-regular MaxCut, at 4 to 10 qubits and 2 to 10 strata, against QNG and Adam.

Critical strengths
  1. The framing in Section 4 is clean. Once the state contracts to the maximally mixed point, every parameter-dependent object lives in the traceless signal, and the metric being quadratic in that signal while the cost gradient is linear is a compact explanation of why the conditioning degrades quadratically in depth rather than linearly. Stating this as a separate failure mode from the barren plateau is worth doing.

  2. The block update is cheap and the experimental grid is broad for its size. Five seeds over four qubit counts, five depths, three problem families, and both noisy and noiseless settings is more than most papers in this area report, and the NAAO numbers do show a consistent ordering in favor of the blocked method.

Critical weaknesses
  1. Proposition 1 is derived from an identity that does not hold for the circuits used to test it. Appendix B.1 writes the parameter derivative as minus i times the generator applied to the output state. That step requires the generator to commute with every gate applied after it. Intra-stratum commutativity in Eq. (11) gives nothing of the kind in QAOA, where the cost stratum is followed by the mixer, or in a hardware-efficient ansatz, where rotations are separated by entanglers. If the step fails, then the vanishing Christoffels, the exact Euler step, and the claim that the retraction defect is identically zero are all unestablished.

  2. The experiments give the blocked method more updates per epoch than the baselines. Algorithm 1 loops over all S strata inside each epoch, so DFS-QNG performs S gradient and metric evaluations per epoch while QNG performs one, and the tables are indexed by epochs with a fixed budget of 30. The reported monotone improvement with the number of strata, for example 0.17 to 0.03 on 4-qubit MaxCut as strata go from one to five, is what more gradient steps would produce even without any geometric benefit.

Questions for authors
  1. Does the generator of a cost-layer ZZ gate commute with the mixer that follows it in your QAOA circuits? If not, which line of Appendix B.1 survives, and can Proposition 1 be restated with the extra hypothesis it actually needs?

  2. Theorem 2 is stated as L tends to infinity with a fixed regularization. But the spike branch requires epsilon much smaller than the squared contraction rate, and for any fixed epsilon that condition fails once the depth is large. Which branch is the theorem about, and does the conclusion depend on epsilon being chosen as a function of depth?

  3. Theorem 1 gives a constant containing the square of the sum of generator norms divided by epsilon, multiplying the square of the observable norm. The derivation in Appendix F ends at four times the observable norm, with no epsilon term. Which expression was proved, and with epsilon small is the rate O(L_R over delta) not vacuous?

  4. What happens to the block update if the commutativity requirement is dropped and the same partition is used with the full block metric? If the gain survives, then the effect comes from inverting small systems rather than from dual flatness. This ablation is absent.

  5. In Tables 1 and 3 the blocked rows at one or two strata are numerically identical to QNG, for example 0.17 against 0.17 and 0.88 against 0.88. Is the claim of no exceptions meant to include ties?

  6. Section 4 supports the spike with a two-qubit plot of gradient norms. Do you have a run where the amplified update increases the loss or destabilizes training, and where rescaling the step size by the inverse contraction factor does not repair it? Without that, how is the spike distinguished from ordinary ill-conditioning?

  7. Equation (10) argues that standard QNG pays a retraction error because it takes a Euler step. At the step sizes used in your experiments, how large is that error relative to the loss decrease per step?

Reviewer_WAcbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies quantum natural gradient (QNG) and Riemannian gradient methods for variational quantum algorithms. QNG preconditions the gradient with the inverse of the Fubini-Study metric pulled back to parameter space, regularized with a Tikhonov term. It is widely considered the theoretically optimal optimizer for variational circuits, but it becomes unstable as circuits get deeper and noisier. The paper aims to explain this instability and to propose a more stable variant.

The first contribution is a convergence analysis. The authors show that the energy is geodesically smooth with respect to the regularized metric, so that Riemannian gradient descent with exact geodesic steps reaches an approximate stationary point at the usual first-order rate. They present this as the first provable convergence rate for Riemannian gradient descent on a variational quantum landscape.

The second contribution is the noise-induced Riemannian spike (NIRS). Under local depolarizing noise with contraction κerr\kappa_{\mathrm{err}} per layer, the metric decays like κerr2L\kappa_{\mathrm{err}}^{2L} with the depth LL, while the Euclidean gradient decays like κerrL\kappa_{\mathrm{err}}^{L} and the Christoffel symbols stay bounded. Inverting the metric then makes the natural gradient grow exponentially with depth when the regularization is weak, and heavy regularization brings back the noise-induced barren plateau. The authors conclude that the instability of noisy QNG has a geometric origin in the metric inversion itself.

The third contribution is Dually Flat Stratified QNG (DFS-QNG). The parameters are split into strata, the paper's term for groups of parameters whose generators commute, such as the layers of hardware-efficient and QAOA circuits. Each update uses only the metric block of one stratum. The authors describe this as block-wise QNG and do not claim the block schedule as new. Their main point is that the metric is constant along a commuting stratum, so that DFS-QNG becomes exact Riemannian block-coordinate descent, with a convergence guarantee and only small metric blocks to invert.

The experiments compare DFS-QNG with QNG and Adam on VQE for the transverse-field Ising model and on QAOA MaxCut, with up to 10 qubits, and only the MaxCut runs on ring graphs include noise. DFS-QNG has the lowest normalized area above optimal (NAAO) in every setting, apart from a few ties with QNG, and QNG gets worse as the number of strata grows. A 2-qubit experiment in the appendix shows the Euclidean gradient shrinking and the Riemannian gradient growing with depth and noise.

Critical strengths
  1. A useful observation about QNG under noise The noise-induced barren plateau results of Wang et al. (2021) concern the Euclidean gradient, but QNG also inverts the noisy metric. The paper points out that the metric is quadratic in the derivatives of the state and the gradient is linear in them, so the two shrink at different rates with depth. This simple observation explains why QNG on noisy circuits needs a step size or regularization that depends on depth. I think QML researchers would take this away from the paper, even though I disagree with how it is framed.

  2. Consistent gains over QNG DFS-QNG only inverts small metric blocks, yet its NAAO is lower than that of Adam in every setting and never higher than that of QNG. The gap to QNG also grows with the number of strata. Even if the method is close to existing work, this is useful evidence that updates with small commuting blocks are much more stable than full QNG in deeper circuits.

Critical weaknesses

I have two main concerns, one about DFS-QNG and one about NIRS.

  1. The insight behind DFS-QNG does not hold up

As the paper says, DFS-QNG "is simply block-wise (or layer-wise) QNG". With all strata evaluated at the same point, as in Algorithm 1, it is the block-diagonal QNG of Stokes et al. (2020), where "all parametric gates within a given layer are assumed to commute". This is also the default approximation in PennyLane's QNG optimizer. The paper mentions such approximations only in passing and does not compare with block-diagonal QNG, so the contribution rests on the claim that commuting strata make QNG exact Riemannian block-coordinate descent.

This claim does not hold. Each update is exact for the metric restricted to its stratum, but the strata are not totally geodesic submanifolds, although the proof of Theorem 3 treats them as such. An update therefore does not follow a geodesic of the full metric. For the one-qubit state RZ(ϕ)RX(θ)∣0⟩R_Z(\phi)R_X(\theta)|0\rangle, a small version of the alternating strata used for VQE, updating ϕ\phi alone moves the state along a circle of latitude, which is not a geodesic unless θ=π/2\theta=\pi/2.

Exact geodesics are also not needed for the rate. Section 5 motivates them by saying that convergence guarantees need a retraction that is accurate to second order, but first-order guarantees hold for general retractions (Boumal et al., 2019). With the Hessian bound of Lemma 8, full QNG with Euler steps and a suitable step size already has the rate of Corollary 1, with a better constant, so commutativity plays no role in the rate of Theorem 3. Theorem 1 and Corollary 1 follow from standard arguments, and I do not see them as a significant contribution on their own.

With the reported experiments, the geometric explanation cannot be separated from a simpler one. At a shared step size, full QNG can take much longer steps than any block method, since the smallest eigenvalue of the full metric is never larger than that of its diagonal blocks. This holds for commuting and non-commuting blocks alike, and it fits the fact that QNG also degrades with depth in the noiseless runs. The paper does not report the step size, regularization, initialization or noise rate, or whether they were tuned for each method.

I would have liked to see QNG with its own tuned step size and regularization, block-diagonal QNG, and DFS-QNG with non-commuting blocks of the same sizes before attributing the gains to "following block-wise accurate geodesics".

  1. NIRS is not established as a failure mechanism

The metric on mixed states in Theorem 2 is never defined. Eq. (2) is the pure-state Fubini-Study metric, while the proof uses a Hilbert-Schmidt representation, and the spike depends on this choice. Under global depolarizing noise of strength pp, for example, the Hilbert-Schmidt metric shrinks by (1−p)2(1-p)^2 and the gradient by 1−p1-p, so the natural gradient keeps its direction and becomes 1/(1−p)1/(1-p) times longer. This is the NIRS scaling, and a step size proportional to 1−p1-p removes it. With the quantum Fisher information, which Koczor & Benjamin (2022) use for noisy QNG, the natural gradient hardly changes unless the state is almost maximally mixed.

As far as I can tell, Theorem 2 therefore shows that the overall scale of one particular noisy metric shrinks faster than the gradient, which is a question of how the step size is chosen. Statement (i) fits this reading, since the Christoffel symbols do not change when the metric is multiplied by a constant. I do not think this supports the conclusion that "metric inversion, not the particular optimization rule, is ultimately responsible for the exponential amplification".

The proof also has gaps. The upper bounds need the smallest eigenvalue of the noisy metric to be at least of order κerr2L\kappa_{\mathrm{err}}^{2L}, but only the largest eigenvalue is bounded, and the metric is singular for redundant parameterizations. The lower bound assumes that the gradient decays no faster than its noise-induced barren plateau bound. Finally, for a fixed regularization ϵ\epsilon the natural gradient is at most 1/ϵ1/\epsilon times the gradient, so the growth stops around the depth where κerr2L\kappa_{\mathrm{err}}^{2L} reaches ϵ\epsilon. The claim that the natural gradient "diverges exponentially with circuit depth" therefore needs ϵ\epsilon to shrink exponentially with depth.

The only experiment for NIRS is a 2-qubit VQE in the appendix with the unregularized metric, and the main experiments in which QNG degrades with depth are mostly noiseless. Since NIRS is the first result in the abstract and the basis for the claim that the instability "has a fundamental geometric origin", these problems affect the core message of the paper.

Questions for authors

Clarification requests

  1. Is DFS-QNG run with all strata evaluated at the same point, as in Algorithm 1, or with the strata updated one after the other, as in the proof of Theorem 3? Does the QNG baseline use the full metric or the block-diagonal approximation that PennyLane uses by default?
  2. Which step size, regularization, noise rate and initialization are used, how many parameters does each stratum have, and were the step size and regularization tuned separately for each method?
  3. Which metric on mixed states is used in Theorem 2, in the 2-qubit appendix experiment and in the noisy MaxCut runs? If the stratum metric is estimated from expectation values of Pauli generators, it tends to a multiple of the identity rather than to zero as the noise grows, so it is not clear to me that the experiments use the metric of Theorem 2.

Suggestions for future work

  1. It would be useful to compare DFS-QNG with block-diagonal QNG, with full QNG with its own tuned step size and regularization, and with the same method on non-commuting blocks of the same sizes.
  2. It would also help to redo the analysis of Theorem 2 with the quantum Fisher information and a normalized natural gradient step. The two-sided bound needs a lower bound on the smallest nonzero eigenvalue of the noisy metric, or Theorem 2 should be stated without it.

Comments on the paper

1 comment
Area_Chair_zcN4OpenReview ↗

Discussion?

Dear authors,

if you want to engage in a discussion, the time is now. The forum will close to the authors on November 18, 2026: https://iclr.cc/Conferences/2027/Dates

Without a rebuttal, there the reviewers are unlikely to change their opinion, which would lead to rejection, most likely.

Sincerely Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.